Stacking transformer layers to create large models results in better accuracies, few-shot learning capabilities, and even near-human emergent abilities on a wide range of language tasks.